What Goes with Red and Blue? Assessing Partisan Cognition Through Conjoint Classification Experiments Stephen N. Goggin John A. Henderson Alexander G. Theodoridis Ph.D. Candidate Assistant Professor Assistant Professor Dept. of Political Science Dept. of Political Science Dept. of Political Science UC Berkeley Yale University UC Merced March 3, 2016 Abstract Political parties can provide valuable information to voters by cultivating distinct associations between their labels, issue priorities, policies and group traits. Yet, there is considerable debate over which associations voters incorporate, and whether these are accurate. In this study, we develop a novel conjoint classification experiment designed to map voters partisan associative networks. We ask respondents to guess the party and ideology of hypothetical candidates given randomized issue priorities and biographical details. Notably, this inferential approach minimizes the biasing effects of partisan boosting in measuring the relative associations voters make between attributes and parties, and the impact these mappings have on candidate evaluations. We find voters consistently link many issues with party and ideological labels, but agree far less on associations with candidate attributes. Our study highlights important heterogeneity in the information value of party reputations, with implications for theories of democratic competence and empirical findings emerging from candidate-vignette designs. For valuable comments we thank Doug Ahler, Henry Brady, Jack Citrin, Jacob Hacker, Greg Huber, Jeff Jenkins, Travis Johnston, Katherine Krimmel, Steve Nicholson, Jas Sekhon, Eric Schickler, Kim Twist, and Rob Van Houweling, as well as workshop participants at the University of Virginia, the CCES Sundance Conference and the ISPS Summer Workshop. All errors are our responsibility. This work was funded by generous research support from the University of California, Merced and Yale University. Replication data happily provided upon request. <goggin@berkeley.edu>, http://www.sgoggin.org/, Institute of Governmental Studies, University of California, 109 Moses Hall, Berkeley, CA 94720-2370 <john.henderson@yale.edu>, http://jahenderson.com/, Institution for Social and Policy Studies, Yale University, 77 Prospect Street, New Haven, CT 06520-8209 <atheodoridis@ucmerced.edu>, http://www.alexandertheodoridis.com, Department of Political Science, University of California, 5200 North Lake Rd., Merced, CA 95343
1 1 Introduction The contextual grasp of standard political belief systems fades out very rapidly... Increasingly, simpler forms of information about what goes with what (or even information about the simple identity of objects) turn up missing (Converse 1964, p. 213). Partisan stereotypes are rich cognitive categories, containing not only policy information but group alliances, trait judgments, specific examples of group members, and performance assessments (Rahn 1993, p. 474). Since at least Converse (1964), scholars have sought to understand what goes with what in the minds of voters, and especially what cluster of things people associate with the parties and political ideologies (e.g., Ansolabehere and Jones 2010; Conover and Feldman 1984; Rahn 1993; Lodge and Hamill 1986). By cultivating reputational associations, parties may provide their candidates with important electoral resources, like low-cost policy signals or issue ownership advantages (Butler and Powell 2014; Petrocik 1996; Pope and Woon 2009; Sniderman and Stiglitz 2012). When reasonably accurate, such associations can give voters cheap information about competing candidates or complex policy proposals, facilitating a minimal form of democratic competence (Dancey and Sheagley 2013; Lupia and McCubbins 1998). More broadly, the transmission of these party stereotypes is also thought to play a role in attitude formation and change (Carsey and Layman 2006; Nicholson 2011; Zaller 1992), potentially mediating constraint or consistency in mass opinion (Converse 1964; Levendusky 2010). While there is general agreement that voters possess some cognitive mapping of the parties, surprisingly little is known about what heuristic information voters incorporate, and how much partisanship or political interest mediates this process. Extensive prior research has considered whether party labels convey information to voters about the policy positions of the parties or candidates (Ansolabehere and Jones 2010; Dancey and Sheagley 2013; Feldman and Conover 1983), the groups, traits or demographics of those identifying with or voting for the parties (Cutler 2002; Petrocik 1996; Rahn 1993), the
2 issue priorities and candidate features of party politicians (Egan 2013; Hayes 2005), and ideological labels or orientations (Sniderman and Stiglitz 2012). Yet, other work finds that many voters possess few, shallow or biased mental images of the parties (Ahler and Sood 2015; Bullock et al. 2015; Converse 1964; Dancey and Sheagley 2016; Kuklinski et al. 2000; Levendusky 2009). Many voters, especially those with limited information, are often unaware of the many issue positions or priorities championed by each party (Levendusky 2009; Zaller 1992). Further, partisan rooting interest can significantly bias beliefs or motivate expressive responses to survey questions that may fundamentally distort measures of voters party stereotypes (Ahler and Sood 2015; Arceneaux 2009; Berinsky 2012; Bullock et al. 2015; Dancey and Sheagley 2016; Einstein and Glick 2013; Hartman and Newmark 2012; Slothuus and de Vreese 2010; Goggin and Theodoridis 2014; Theodoridis 2012). Our aim in this study is threefold. First, we seek to clarify the cognitive role partisan and ideological associations play in voters minds. We synthesize findings across research on party stereotyping to develop what we call the partisan associative network, the many-dimensional cluster of durable associations that cognitively cohere with partisan or ideological objects, and that come to mind when features in the network are referenced. Such cognitive associations have been the focus of disparate literatures, with many existing theoretical conceptions and empirical findings standing in conflict and unresolved. In bridging this research, we highlight two potential sources of disagreement, party identity (PID) and political knowledge, both of which are thought to mediate the formation and activation of associative networks. Second, we develop and implement a novel experimental design to robustly measure voters associative networks, minimizing the influence of partisan bias resulting from voters PID. To do this, we call upon the conjoint experimental framework pioneered in political science by Hainmueller et al. (2014); Hainmueller and Hopkins (2015). We ask respondents to guess the party or ideology of fictional candidates given a set of
3 randomly generated issues and candidate attributes. The method allows us to measure both the partisan direction and relative strength of associations across a wide variety of issues and candidate traits, as well as to assess how respondents political knowledge and PID mediates these associative networks. We implement this design fielding experiments in two modules of the 2014 Cooperative Congressional Election Survey (CCES). From the analysis, we find that issue associations largely mirror expectations from work on issue ownership, though interesting exceptions emerge. We find remarkably similar results in analyzing ideology rather than party guessing, as well as when stratifying associations by party identification (PID), providing strong evidence that both Democrats and Republicans agree about which issues go with which parties and ideological orientations. Yet, in breaking down candidate evaluations by PID we see clear evidence of polarization in affective orientations, bolstering concerns that partisan rooting interest could be introducing a source of bias in previous studies of party stereotyping. We also find that voters associate a number of candidate traits with the parties, though some are more consistently linked (e.g., gender, religion, occupation) than others (e.g., family background, military service). In line with expectations from research on political interest (e.g., Converse 1964; Levendusky 2009; Zaller 1992), we find clearer ideological and partisan images for those with higher knowledge, reflecting a better understanding of how issues, candidates and parties go together. Lastly, we examine whether the party and ideological associations present in voters minds resemble the parties actual priorities and coalitions. A great deal of research has studied, with somewhat mixed results, whether politicians prioritize issues in Congress (Egan 2013; Woon and Pope 2008) or elections (Damore 2004; Sides 2006) consistent with measures of voter stereotyping. Much less work has investigated whether voters accurately incorporate the associative information being projected by parties. We find little evidence that information emerging from election competition is driving the party associations we observe, as candidates systematically deviate from prioritizing issues as
4 expected by voters. Regarding candidates personal attributes, we find voters rarely hold inaccurate stereotypes, though they do often miss meaningful associations that are present. This study provides perhaps the best evidence to date about the mental images voters carry around of the parties and their ideological labels. As such, our findings have a number of implications for work on democratic competence, representation and political communication. We find stronger associations between issues than traits, indicating that voters are more aware of differences in the parties policy priorities, rather than in the descriptive composition of their standard bearers. A possible explanation is that parties avidly seek to differentiate their legislative priorities, while simultaneously aiming to expand their demographic coalition. More generally, the presence (and absence) of associations across particular domains can provide greater insight into the way voters incorporate and use political information, as well as how party reputations facilitate this effort. Our findings also help clarify conditions under which the party brands may constrain candidate behavior, and thus improve electoral accountability. Finally, our study makes a significant contribution to experimental design by separating association from evaluation. In utilizing quiz-like inferences rather than direct survey instruments, our approach minimizes one of the core pitfalls in previous efforts to examine cognitive maps of partisan space, namely the tendency of partisan identifiers to overweight positive associations with their favored party s candidates. We suggest caution in interpreting the results from candidate vignettes or other studies that prime partycorrelated information, and recommend the use of similar quiz-like designs to minimize partisan boosting when crafting future experimental analyses. 2 Understanding Partisan and Ideological Associations Most prior work on party stereotyping is rooted in a basic idea: voters possess certain mental images of the parties, which they use to evaluate parties, policies or candidates
5 in elections (Arceneaux 2009; Feldman and Conover 1983; Nicholson and Segura 2012; Rahn 1993; Sniderman and Stiglitz 2012; Snyder and Ting 2002; Woon and Pope 2008). Fundamental to the partisan stereotyping process, these mental images can be signaled simply through reference to the party label. This notion of stereotyping underpins many important theoretical findings in political science, including how parties manage elite ambition (Fiorina 1980; Snyder and Ting 2002) and facilitate electoral competition (Petrocik 1996; Sniderman and Stiglitz 2012), as well as how voters incorporate and use heuristic information (Lee 2009; Lupia and McCubbins 1998), or come to form and change opinions on policies (Carsey and Layman 2006; Levendusky 2010; Nicholson 2011). While many of these theories consider party stereotypes as positive public goods, others explore more problematic implications of party reputations, including biases in voter decision-making (Rahn 1993), learning (Dancey and Sheagley 2013), and information recall (Coronel et al. 2014), as well as broader concerns about democratic accountability (Arceneaux 2009; Damore 2004). 2.1 Connecting Ownership Theories to the Party Policy Brands Though scholars generally agree that voters engage in some stereotyping, there is considerable disagreement about the information that voters incorporate, or how that information relates to their choices. One important debate is over the extent to which party labels clarify the ideological or positional differences between parties and candidates (Fiorina 1980; Levendusky 2010; Lupia and McCubbins 1998; Snyder and Ting 2002; Sniderman and Stiglitz 2012), or serve mostly as non-policy or valence heuristics for competence, good government or policy success (Cox and McCubbins 2005; Lee 2009; Petrocik 1996). Notably, each has important consequences for how voters would balance rewards for legislative unity against enacting undesirable policies (e.g., Cox and McCubbins 2005; Egan 2013; Lee 2009). Yet, it has proven difficult to empirically adjudicate between the two accounts (e.g., Butler and Powell 2014).
6 A part of the challenge is in determining how voters weigh competing types of reputational information when evaluating candidates. Most valence accounts, including party ownership, see voters as indifferent to the policy orientations of politicians, so long as they do something to resolve problems arising when governing (Egan 2013). Yet, with polarization in Congress, the parties reputations appear to be increasingly grounded in their persistent divergence on policies, and around policy successes that divide electorates. As consequence, this could lead voters to link particular ideological labels to the parties, rather than just issue priorities, and to use those labels when choosing candidates (Sniderman and Stiglitz 2012; Snyder and Ting 2002). However, voters may not incorporate such positional reputations, especially when these require additional information about how particular issues or policy proposals map on to ideological conflict (Butler and Powell 2014). Indeed, many voters mistake which party is the more liberal on a range of policy items, and overall (Levendusky 2010; Sniderman and Stiglitz 2012). This may lead less informed voters to focus on keeping score of legislative wins rather on the specific policy details being fought over (Egan 2013). On the other hand, since parties, policies and ideologies tend to cluster, candidates may be able to signal some ideological information simply by highlighting issue priorities (Henderson 2015a). A particularly influential valence account of party stereotyping is the theory of party ownership. In traditional form, articulated most clearly by Petrocik (1996), ownership derives from voters overall beliefs that some parties do better handling certain issues when in office than do their partisan competitors. More recently, scholars have incorporated candidate traits and biographies alongside issues into the basic framework, arguing parties have cornered certain candidate-level qualities alongside their issue portfolios (Goggin and Theodoridis 2014; Hayes 2005, 2010). Accordingly, parties have incentives to cultivate such impressions, so that, once in place, candidates can use them to win elections. By defining elections to be about the issues owned by that candidate s party, candidates remind voters that the most important issues of the day are the ones that she and her
7 team are better equipped to handle or care more about in office. Further, certain traits may lead voters to see candidates as more competent in handling related issues, may be desirable on their own, or may resonate with certain subsets of voters. By coming to own particular traits, parties can further enhance the credibility or electability of their candidates (Hayes 2005, 2010). Within ownership accounts there is considerable ambiguity over whether trait ownership is merely a byproduct of parties owning issues, or the reverse. For example, parties trusted to resolve national defense crises may also be viewed as more likely to demonstrate strong leadership abilities. Yet, the opposite causal direction is also clearly plausible: parties that draw candidates with military backgrounds of leadership may be viewed as better able to handle issues of national defense. Measuring how voters evaluate candidates given variation in traits and issue priorities cannot clarify whether either type of ownership independently influences voter behavior, since biographical and policy signals are likely to be correlated in voters minds. Ultimately, what is needed is a way to decouple these signals when measuring party associations and candidate evaluations. Not surprisingly, scholars also disagree about where exactly party stereotypes originate. According to ownership accounts, parties cultivate reputations by repeatedly addressing or resolving problems in particular issue-areas (Egan 2013; Petrocik 1996). Yet, other scholars see party reputations as emerging specifically from legislative enactments in Congress (Butler and Powell 2014; Cox and McCubbins 2005; Woon and Pope 2008), ideological screening in primaries (Snyder and Ting 2002), the party platforms or party coalition maintenance (Karol 2009), or even the biographical traits of candidates (Hayes 2005). Naturally, each source implies very different information is being signaled to and incorporated by voters in the party label.
8 2.2 Defining Party Associative Networks In this study, we aim to clarify what information voters incorporate into their mental images of the parties, and whether such information systematically varies across the electorate. To this end, we synthesize a number of findings from prior research on party stereotyping into what we call party associative networks. By associative networks, we mean the multi-dimensional cluster of political information that tends to cohere at a cognitive level with other related objects in voters minds. We focus primarily on the content of these networks, which includes the issues, positions, biographical attributes, and personality traits that get associated with the party labels, or affiliated ideologies. Most prior research on party stereotypes studies the consequences of activating particular associations in the political world. In contrast, our focus is on conceptualizing broader characteristics of this network as a whole, including how such durable associations form and relate to each other, as well as interact with voter partisanship or political interest. In broad terms, party associative networks define what goes with what in voters minds when thinking about the parties. This concept is based off of Converse s (1964) work on belief systems, which offers a framework for considering the presence, strength and consistency of associations between idea elements. Using a variety of survey techniques, Converse (1964) examines the extent to which voters grasp how ideas fit together, and are unconstrained in their own thinking, both at a given moment and longitudinally. Of course his major finding is that voters are largely not able to recall or recognize affiliated attitudes between parties, candidates, issues and ideologies. Here we extend what Converse calls static constraint into thinking about the cluster of political information voters map together: In the static case, constraint may be taken to mean the success we would have in predicting, given initial knowledge that an individual holds a specific attitude, that he holds certain further ideas and attitudes (Converse 1964, p. 207). In linking this constraint to information rather than opinion, we aim to exploit the predictive value particular dimensions have on partisanship as a way to estimate the relative
9 weight of that association on voters minds when the party label is primed in political context. A major concern to Converse (1964), and much of the subsequent research on stereotyping, is that voters may selectively access certain kinds of information in ways that distort the mental images they possess and utilize. Numerous scholars have found that political interest significantly mediates the amount and type of information voters receive or accept (Converse 1964; Levendusky 2010; Lupia and McCubbins 1998; Zaller 1992). A number of important findings, including Converse s (1964) seminal study, suggest that many voters possess rather shallow or inconsistent policy attitudes, attributable to their limited knowledge, interest or experience in politics. Exposure to political information then may be a predicate to forming stereotypes about the parties, which themselves are necessary to form or change opinions on policies (Carsey and Layman 2006; Levendusky 2010; Zaller 1992). By implication, constraint in voter opinion may significantly depend on the combination of clear policy signals being sent by party elites and sufficient voter interest to receive and incorporate these signals into beliefs about the parties. In the extreme, this raises the possibility that politically uninterested voters may even be uninformed about the kinds of stereotypical information they would most require to effectively use the party label when voting. Beyond political engagement, rooting interest stemming from party identification can also significantly mediate the mental images voters form about parties. Similar to political interest, partisanship may serve as an information screen yielding selectivity in the kinds of sources voters seek out or avoid (Arceneaux and Johnson 2013; Henderson and Theodoridis 2016). More powerfully, however, partisanship can distort the images voters possess about the demographic traits of party voters or candidates (Ahler and Sood 2015; Goggin and Theodoridis 2014), or their policy attitudes and ideological extremity (Ahler 2014; Rahn 1993; Feldman and Conover 1983). Party identification can bias how people use issue primes or other party heuristic information (Arceneaux 2009; Dancey and Shea-
10 gley 2016; Slothuus and de Vreese 2010). Finally, partisanship can even lead voters to be susceptible to rumors or misinformation about out-party politicians, reflecting a mix of either meaningful beliefs (Einstein and Glick 2013; Hartman and Newmark 2012) or expressive attitudes (Berinsky 2012; Bullock et al. 2015). Crucially, partisanship can distort not only the images voter possess, but also scholars abilities to measure voter stereotypes. For example, the standard measure of issue ownership asks voters which party they think would do a better job handling a number of specific issues. Responses are usually interpreted as tapping unbiased evaluations of the parties abilities. However, there is an obvious concern that many partisans will (expressively or genuinely) state they trust their own party to handle every issue (Egan 2013; Goggin and Theodoridis 2014). Independent identifiers in contrast, though unbiased, may be less informed about the parties priorities or performance in office. In the extreme case, partisans responses could cancel out, so that relatively uninformed independents tell us which issues the parties handle better in office. This concern is not unique to issue ownership, and is likely to emerge in many other measures of party stereotyping. As consequence, many previous findings about party stereotypes could reflect the mental images of the least rather than the most informed voters, or could be distorted by the way motivated partisans respond to survey items. In the experiments presented below, we aim to add some conceptual and empirical clarity to the examination of party brands. We do so by mapping which traits and issues get linked to the parties, and assessing how these linkages influence voter evaluations of particular candidates who exemplify those features. The conjoint experimental task we develop allows us to consider a large number of features simultaneously. In this study, we examine two particular dimensions in the associative network: issue priorities and candidate traits. This focus is motivated by major theoretical findings and challenges in work on both ownership and party policy brands, including distinguishing the independent influence of traits and issues, as well as whether these carry any ideological or
11 positional information. In contrast to these literatures on the consequences of stereotyping, however, we aim to investigate the extent to which the central associations advanced in this work are present in voters minds in order to function as hypothesized. 3 Developing a Robust Measure of Cognitive Associations To provide a robust measure of party associative networks, we utilize a novel conjoint experimental design (Hainmueller et al. 2014). In our application, we ask respondents to infer the party and ideology of fictional candidates given a set of issues and candidate attributes randomly presented to them. Notably, the method allows us to measure both the direction and strength of associations across a wide variety of issues and traits, as well as to assess how respondents characteristics influence these associations. Conjoint studies, long used in product marketing research, have recently been introduced in political science (Hainmueller et al. 2014; Hainmueller and Hopkins 2015). This method is especially well-suited to our analysis, as it allows for manipulation of multiple attributes within a variety of factors. Our study differs in significant ways from previous applications of the conjoint design in that we do not use a paired comparison. Rather, in each task, respondents are asked to assign a single fictional candidate to one of two categories - either the two parties (Democrat or Republican), or the two ideological labels (Liberal or Conservative). Paired comparisons, while common in conjoint studies, are not an essential feature of the method. In fact, the single target actually reduces the number of assumptions necessary for analysis of the data. 1 In our experimental frame, respondents are shown a set of hypothetical candidate 1 Inferences from conjoint data require a few important assumptions. Most notably, we assume stability of evaluations, that there are no carry-over effects, and that there are no profile-order effects (Hainmueller et al. 2014). These assumptions are necessary because we gain statistical power by having each respondent evaluate four different hypothetical candidates. In this case, we believe a paired comparison would have proven unwieldy for
12 characteristics, presented in a table format that includes the randomly manipulated candidate attributes drawn from a number of trait and issue dimensions. Figure 1 shows both the introductory page seen by respondents prior to the task and a sample candidate categorization page. In prefacing the experiment, we depict the candidate information as coming from a questionnaire to increase the verisimilitude of the task. As seen in Figure 1, the left side of the table shown to respondents indicates the category of attribute information requested in our fictitious questionnaire (e.g., Gender, 1st Issue Priority), while the right side presents the fictional candidates response to that particular item. The full list of categories and levels in the experiment is shown in Table 1. 2 Each candidate had one of the levels for each factor randomly inserted into each category with equal probability. 3 The order of the factors was randomly assigned (with the three issue priorities always listed together and in numerical order) at the level of the respondent. The wording of the issue priority levels was designed to avoid signaling an issue position or policy direction, while still specifying a priority. For example, Tax reform was used as opposed to language referencing either tax cuts or increases. We examine results from two separate conjoint experiments. The first of these was fielded in two pre-election survey modules (Henderson 2015b; Theodoridis 2015) of the respondents, especially given the binary nature of our primary outcome measure. This is partly because, unlike the paired choices in other conjoint tasks (e.g. selecting which immigrant should be admitted or which candidate deserves your vote), there is no realworld analogue for guessing which candidate is more likely a Democrat or Republican. 2 The factors were selected to provide a reasonably detailed candidate description and based upon characteristics often linked to political outcomes. A notable omission is race. We elected to remove race for fear that the anticipated massive Democratic effect for black candidates would mask associations on other dimensions. 3 Each candidate had a 1st Issue Priority, a 2nd Issue Priority, and a 3rd Issue Priority, with these issues sampled without replacement from the list in Table 1.
13 2014 Cooperative Congressional Election Study (CCES) (Ansolabehere 2015). 4 In this experiment, we ask respondents to guess whether each presented candidate was a Democrat or Republican, rate how sure they were about their guess, and evaluate the candidate from very unfavorable to very favorable on a 11-point scale. A total of 2000 respondents completed this task four times during the survey, resulting in 8000 observations. 5 The second conjoint experiment, presented in the post-election portion of Henderson (2015b), was identical in design, except respondents were asked to guess the candidate s ideology (liberal or conservative) rather than the candidate s party. This conjoint experiment was completed by a total of 1000 respondents, each completing the task four times, resulting in 4000 observations. 6 For all analyses, standard errors are clustered at the level of the respondent. In asking respondents to guess party or ideology given various features, we provide some of the first empirical evidence (rather than making the assumption) that certain issue priorities and candidates traits are in fact associated with different parties. Further, the quiz-like feature of guessing minimizes the possibility that partisan bias is producing the recovered associations between features and parties (Henderson 2015a). In this vein, 4 The CCES is fielded online by YouGov in the weeks just prior to and just after November 2014 s Election Day. The analysis in this paper does not use sampling weights. This is done to avoid the risk of post-hoc weights exacerbating balance issues across experimental conditions or impacting estimates within cells, something of particular concern with a conjoint design. 5 Within this sample, 48.7% of respondents were Democratic or Democratic leaners, 13.4% were pure independents, and 33.0% were Republican or Republican leaners, with 5.0% listing other or declining to answer. 6 Within this sample, 46.9% of respondents were Democratic or Democratic leaners, 15.3% were pure independents, and 33.0% were Republican or Republican leaners, with 4.8% listing other or declining to answer.
14 (a) Intro (b) Conjoint Task Figure 1: Conjoint Task: This is how the conjoint task was presented to respondents.
15 Table 1: Experimentally Manipulated Factors and Levels Factor Gender Family Religion Military Occupation 1st, 2nd and 3rd Issue Priorities (without replacement) Levels Male Female Unmarried Married Married with one son Married with one daughter Married with one daughter and one son Married with three sons Married with three daughters Catholic Jewish Mainline Protestant Evangelical Protestant None Listed None National Guard US Army Major in US Army Small Business Owner Attorney Doctor CEO Farmer Teacher Factory Foreman Construction Contractor Political Staffer Retail Manager Strengthening national defense Preventing future terrorist attacks Promoting strong moral values Addressing the immigration problem Tax reform Fighting against illegal drugs Reducing crime Promoting trade with other nations Reducing the budget deficit Creating new jobs Strengthening the economy Promoting energy independence Improving education Strengthening Social Security Preserving Medicare Protecting the environment Improving health care Assistance to the poor and needy
16 the subsequent evaluations frame can actually show how serious rooting interest may be in leading voters to make different evaluations about candidates in observing different issue priorities and traits. Finally, in quizzing about ideological as well as party labels, we can investigate more general theoretical views of party reputations, and in particular, whether voters do perceive certain issues and traits as informative about the policy orientations of candidates. 4 Mapping Partisan and Ideological Associations 4.1 Associations with Party The main results of our first conjoint task asking respondents to guess the party of a hypothetical candidate are shown in Figure 2. The Figure displays the marginal effect each attribute had on the probability a respondent classified the candidate as a Republican, relative to the omitted level of each experimental factor. 7 Within each factor, the omitted variable is displayed first, with no 95% confidence interval. It is important to note that estimated effects are relative to the omitted categories. If we omit a different level for each category, while the direction and appearance of the plot may change, the relative distance between levels will not change. Several important patterns among the personal attributes emerge in Figure 2. First, female candidates were guessed to be more likely Democratic candidates. Second, familial status - one s marital status and children - appears to have little effect on the guessed party, even as the gender and quantity of children varies significantly. Third, both Mainline and Evangelical Protestants were guessed to be more likely Republican, with Evangelical even significantly more Republican than Mainline. Fourth, while mili- 7 Because the classification task is binary and no don t know option was given, a negative effect signals a respondent was more likely to classify the candidate as a Democrat in the presence of that attribute.
17 Experimentally Manipulated Variable Gender Male Gender Female Family Unmarried Family Married Family Married with one son Family Married with one daughter Family Married with one daughter and one son Family Married with three sons Family Married with three daughters Religion None Listed Religion Catholic Religion Jewish Religion Mainline Protestant Religion Evangelical Protestant Military None Military National Guard Military US Army Military Major in US Army Occupation Retail Manager Occupation Small Business Owner Occupation Attorney Occupation Doctor Occupation CEO Occupation Farmer Occupation Teacher Occupation Factory Foreman Occupation Construction Contractor Occupation Political Staffer Issue Strengthening the economy Issue Strengthening national defense Issue Preventing future terrorist attacks Issue Promoting strong moral values Issue Addressing the immigration problem Issue Tax reform Issue Fighting against illegal drugs Issue Reducing crime Issue Promoting trade with other nations Issue Reducing the budget deficit Issue Creating new jobs Issue Promoting energy independence Issue Improving education Issue Strengthening Social Security Issue Preserving Medicare Issue Protecting the environment Issue Improving health care Issue Assistance to the poor and needy 0.2 0.0 0.2 Change Pr(Guess Candidate is Republican) Figure 2: Guessing Candidate s Party: Estimates are OLS regressing party guesses on all factor levels. Standard errors are clustered on the respondent, with error bars displaying 95% confidence intervals. Estimates with no error bars are the excluded levels of each experimental factor. Issues are coded as present if they were in either the first, second, or third issue priority for the candidate. All variables are coded 0-1, with the dependent variable coded as 1=Republican, 0=Democrat.
18 tary service trends toward guessing Republican in all conditions, it is the strongest signal of being a Republican when paired with officer status. Finally, several important patterns regarding the occupational background of the candidate emerge. Relatively high-status backgrounds - attorney, CEO, doctor - have little effect on the candidate s guessed party. However, small business owner leads to more Republican guesses, while teacher and factory foreman lead to more Democratic guesses. Several other occupations - farmer, construction contractor, and political staffer - all have negligible effects as well. Figure 2 also provides important evidence regarding the brand content of issues for both parties. Candidates presented to respondents noted three issue priorities - the figure displays the effect of having an issue listed in any of the three positions. 8 First, when the candidate highlighted national defense, preventing terrorist attacks, promoting moral values, and reducing the budget deficit, respondents were all more likely to guess the candidate was Republican. Second, when the candidate highlighted education, Social Security, Medicare, the environment, health care, and assistance to the poor and needy, respondents were more likely to guess the candidate was Democratic. Third, immigration, tax reform, drugs, crime, trade, creating jobs, and energy independence were all not significant signals of partisanship, when compared to the baseline (strengthening the economy). Some of these are surprising given expectations regarding owned issues. The above results largely comport with existing theoretical accounts of party stereotyping, especially for issue priorities, but also for some personal attributes. However, it is unclear whether these cognitive party associations are being inflated or masked by respondents partisan identities. For this reason, we break the effects apart by respondent 8 Separating the issues by 1st priority, 2nd priority, and 3rd priority leads to substantially the same conclusions. Issues were much stronger signals (leading to more significant results and smaller confidence intervals) when they were higher priority, but the direction and relative magnitude of the effects were very similar across the three positions. For clarity of presentation, we collapse across positions here.
19 PID in Figure 3(a). 9 As shown in Figure 3(a), Republicans and Democrats infer similar partisan signals coming from many of the candidate trait dimensions. Perhaps the most notable of these is candidate gender, where female politicians are much more likely to be inferred as Democrats by both Democratic and Republican identifiers. We find similar agreement for military service and occupation attributes. 10 In contrast, we find that partisans strongly disagree in their inferences given variation in family status. This is likely because each family category is viewed positively in voters minds relative to being unmarried, so that partisans are attributing these positively valenced characteristics more to in-party, rather than out-party candidates. Interestingly, the reverse is true for the religious cues, including Catholic, Jewish, and Evangelical Protestant. For all these, respondents of one party are more likely to guess the cue belongs to a partisan of the other party. However, these effects do not always overwhelm the main party signal - both Republicans and Democrats rate Evangelical Protestants as more likely to be Republican. With respect to issues, there is generally strong agreement about which issue priorities Democratic and Republican identifiers associate with each party. Moreover, similar to the findings in Figure 2, these associations largely resemble core predictions emerging from issue ownership accounts. Assistance to the poor, improving health care or education, 9 We exclude independent respondents for clarity in presentation. In virtually all cases, the estimate for independents lies between that for Democrats and Republicans, though with relatively wide confidence intervals given their smaller proportion of the survey sample. 10 Within the occupation factor, we see some positive and negative effects of partisan identification. For instance, small business owner appears to have a positive valence and is claimed by partisans from both sides. However, occupations such as attorney and political staffer have a negative valence and are pushed toward the out-party by partisans.
20 and strengthening or preserving Medicare and Social Security are all expected to be prioritized by Democrats, while strengthening national defense, promoting moral values and preventing terrorist attacks are (mostly) expected to be promoted by Republican politicians. Interestingly, partisan identifiers also largely agree about which issues are contested, that is roughly equally likely (off the economy baseline) to be promoted by Republican and Democratic office-seekers, such as tax reform, reducing crime, and promoting trade, for example. However, there are a few issues where partisans, though agreeing on party direction, disagree on the magnitude of the associations they infer. Relative to the beliefs of Democratic identifiers, Republicans appear to believe that Social Security and Medicare are more likely to be prioritized by some Republicans, though both are firmly on the Democratic side of the spectrum. Finally, with respect to the environment, we see a negatively valenced issue priority, with Republicans taking it as a very strong cue of the candidate being Democratic, while Democratic respondents viewing it as a much weaker cue of a candidate being Democratic. All these results highlight several important findings - while there is substantial agreement in many domains on the association of certain personal attributes and issue priorities with party, there is also significant heterogeneity by the partisanship of respondents on many attributes. That is, even when tasked with a relatively objective guessing task, respondent partisanship leads them to be more likely to assign positively valenced attributes as copartisan, while negatively valenced attributes are more likely to be assigned to the other party. However, despite these significant effects of partisan boosting, many of these disagreements among partisan respondents still do not lead to the opposite inference of candidate party. Finally, we can examine the patterns of guessing by the political knowledge of respondents, as shown in Figure 3(b). As it is quite clear for many of the biographical attributes and issues, there is very little disagreement between those with high political knowledge and low political knowledge. Where there is statistically significant disagreement, there is
21 a larger effect for those with high political knowledge. That is, those with more political knowledge are likely to see the association, especially over issue priorities. 4.2 Impact on Candidate Evaluations Immediately after guessing the partisanship of the presented candidate, respondents were asked to evaluate how they felt about the candidate from very unfavorable to very favorable. This item helps us to evaluate the positive or negative valence of each piece of information within each factor, particularly how it may be differential among partisan subgroups of respondents. Figure 4(a) presents the marginal effect of each information item on the evaluation of the candidate, scaled from 0 to 1. Therefore, an effect of 0.05 means a 5% decrease in the evaluation of the candidate, relative to the omitted level of each factor. The results in Figure 4(a) reveal several striking patterns. First, with a few prominent exceptions, many of the pieces of information have relatively little effect on the overall evaluation of the candidate. Notably, gender appears to have little effect on overall evaluation, and with the exception of Evangelical Protestant leading to a more negative evaluation, few religious categories do either. Familial status, while not significant, appears to indicate that married candidates are weakly preferred to unmarried ones. With respect to military service, little overall effect on evaluations emerges, although it appears those with military background are viewed weakly more positively than those with none. Occupational status is relatively mixed, with many of the occupations being evaluated as equivalent to the omitted category, retail manager. Small business owner and factory foreman, however, are both evaluated significantly more positively than many of the other occupations. With respect to issue priorities, it appears that many of the issues make little difference in the evaluation of the candidate. 11 Notably, protecting Social Security 11 The seemingly overall negative trend in evaluations among many of the issues is simply a result of the omitted issue priority being the near universal strengthening the
22 Gender Male Gender Female Family Unmarried Family Married Family Married with one son Family Married with one daughter Family Married with one daughter and one son Family Married with three sons Family Married with three daughters Religion None Listed Religion Catholic Religion Jewish Religion Mainline Protestant Religion Evangelical Protestant Military None Military National Guard Military US Army Military Major in US Army Occupation Retail Manager Occupation Small Business Owner Occupation Attorney Occupation Doctor Occupation CEO Occupation Farmer Occupation Teacher Occupation Factory Foreman Occupation Construction Contractor Occupation Political Staffer Issue Strengthening the economy Issue Strengthening national defense Issue Preventing future terrorist attacks Issue Promoting strong moral values Issue Addressing the immigration problem Issue Tax reform Issue Fighting against illegal drugs Issue Reducing crime Issue Promoting trade with other nations Issue Reducing the budget deficit Issue Creating new jobs Issue Promoting energy independence Issue Improving education Issue Strengthening Social Security Issue Preserving Medicare Issue Protecting the environment Issue Improving health care Issue Assistance to the poor and needy Democratic Respondent Republican Respondent Gender Male Gender Female Family Unmarried Family Married Family Married with one son Family Married with one daughter Family Married with one daughter and one son Family Married with three sons Family Married with three daughters Religion None Listed Religion Catholic Religion Jewish Religion Mainline Protestant Religion Evangelical Protestant Military None Military National Guard Military US Army Military Major in US Army Occupation Retail Manager Occupation Small Business Owner Occupation Attorney Occupation Doctor Occupation CEO Occupation Farmer Occupation Teacher Occupation Factory Foreman Occupation Construction Contractor Occupation Political Staffer Issue Strengthening the economy Issue Strengthening national defense Issue Preventing future terrorist attacks Issue Promoting strong moral values Issue Addressing the immigration problem Issue Tax reform Issue Fighting against illegal drugs Issue Reducing crime Issue Promoting trade with other nations Issue Reducing the budget deficit Issue Creating new jobs Issue Promoting energy independence Issue Improving education Issue Strengthening Social Security Issue Preserving Medicare Issue Protecting the environment Issue Improving health care Issue Assistance to the poor and needy 0.2 0.0 0.2 Change Pr(Guess Candidate is Republican) 0.4 0.2 0.0 0.2 0.4 Change Pr(Guess Candidate is Republican) (a) Guessing Party by Party ID Experimentally Manipulated Variable Experimentally Manipulated Variable High Knowledge Respondent Low Knowledge Respondent (b) Guessing Party by Political Knowledge Figure 3: Guessing Candidate s Party by Respondent Party ID or Political Knowledge: Estimates are OLS regressing party guesses on all factor levels. Standard errors are clustered on the respondent, with error bars displaying 95% confidence intervals. Estimates with no error bars are the excluded levels of each experimental factor. Issues are coded as present if they were in either the first, second, or third issue priority for the candidate. All variables coded 0-1, with the dependent variable coded as 1=Republican, 0=Democrat. Partisans include independent leaners. Pure independents are excluded. Political Knowledge was measured with a 6-item measure on the CCES Common Content, then median split.
23 and preserving Medicare result in significantly higher evaluations of the candidate when listed as priorities than many other issues. Though interesting, these findings may disguise considerable heterogeneity by the partisanship of respondents. As noted in the previous section, respondents of both parties have distinctive views about what attributes are positive or negative in candidates. For this reason, we show the effects of the different pieces of information, broken down by respondent partisan identification, in Figure 4(b). As with the results from the previous section, the most interesting findings here emerge when the evaluations by Democratic and Republican respondents diverge. With respect to personal attributes, a candidate being female results in a positive evaluation among Democratic respondents, while such candidates are viewed quite negatively by Republican respondents. With respect to many of the levels of familial status and religion, there appear to be few differences. Notably, however, a candidate being an Evangelical Protestant yields a negative effect among Democratic respondents and a neutral (but not positive) effect among Republican respondents. On military service, an interesting pattern emerges. Among Democratic respondents, military service results in a significantly less positive evaluation than among Republican respondents. Although many of these estimates are not significantly different from the no military background condition, they are significantly different from each other, with Democratic respondents viewing it as a weakly negative cue, and Republican respondents viewing it as a weakly positive cue. There is little divergence among occupations, with the exception of small business owner, which results in significantly higher evaluations among Republican respondents than among Democratic respondents. Perhaps most striking is the large gaps that emerge in the candidate evaluations by partisan respondents across different issue priorities. Of particular note, candidates that prioritize typically Republican-owned issues (e.g., national defense, preventing terrorist economy, as that is viewed as more positive than the bulk of issue priorities.